MiniMax H3 mmproj support for Qwen3 VL 32B GGUFs

[Nif00]

No description provided.

[allanmeng]

I hit the same problem from the Krea2 side (#464) and traced it to the Qwen3-VL mmproj
mapping. Your branch is the only one that wires qwen3vl into gguf_mmproj_loader at
all — but as written the DeepStack keys can't land, so Krea2 still fails. Two defects:

1. CLIP_VISION_QWEN3_MAP maps "v.deepstast.". The real tensors in
   Qwen3-VL-4B-Instruct-mmproj-bf16.gguf are v.deepstack.{5,11,17}.{norm,fc1,fc2}.{weight,bias}
   (18 tensors, verified with GGUFReader). "deepstast" never matches, so the DeepStack
   tensors pass through unrename­d.

2. The layer-index remap (ck.5.→st.0., ck.16.→st.1., …) is bound to a ck. prefix
   that doesn't exist in the file. Even with (1) fixed, you'd emit
   model.visual.deepstack_merger_list.5/11/17.*, but comfy.sd.detect_te_model() only
   probes model.visual.deepstack_merger_list.0.norm.weight — so it still falls back to
   QWEN3_4B and Krea2 still raises "expects 12x2560=30720 features but got 2560".

I have a small fix (loader.py only, +53/-3) that derives the index by sorting the layer
ids present in the file instead of hardcoding 4B/8B, plus an offline check against
qwen3vl_4b_fp8_scaled.safetensors (315/315 keys, cos ≈ 1.000) and a B580 hardware run.
Happy to rebase it on top of your branch if you'd rather keep everything in this PR —
otherwise I'll open it separately as a focused fix.